A memory and meta learning based solution in graph continual learning

Abstract

Deep learning models have proven to perform successfully at different tasks such as classification and regression. Continual learning (CL) aims for a model to learn various tasks sequentially. However, when the models are expected to adapt to incoming tasks without maintaining their performance on previous tasks, they tend to forget the previous tasks. This phenomenon is called catastrophic forgetting and catastrophic forgetting is the main challenge in the CL area. Catastrophic forgetting refers to the scenario where a model tends to forget the previous tasks it had been trained on and adjusts its parameters to perform the task it is actively being trained on. Since it is inefficient to train multiple models to perform multiple tasks, CL aims to train a single model such that it can perform on multiple tasks without losing information during the training process. In addition to catastrophic forgetting, CL also focuses on capacity saturation which is another challenge focusing on the effects of the model architecture on learning. CL is currently an emerging research field topic. However, the CL studies mainly focus on image data and there is much to discover in CL research focusing on graph-structured data or graph continual learning (GCL). The proposed solutions for GCL are mainly adapted from the general CL solutions, therefore, there is much to discover in GCL field. However, since the graph-structured data has different properties compared to image data, the graph properties need to be considered when GCL is studied. In this thesis, we focus on continual learning on graphs. We devise a technique that combines two uniquely important concepts in machine learning, namely "replay buffer" and "meta learning", aiming to exploit the best of two worlds to successfully achieve continual learning on graph structured data. In this method, the model weights are initially computed by using the current task dataset. Next, the dataset of the current task is merged with the stored samples from the earlier tasks, and the model weights are updated using the combined dataset. This aids in preventing the model weights converging to the optimal parameters of the current task and enables the preservation of information from earlier tasks. We choose to adapt our technique to graph data structure and the task of node classification on graphs and introduce our method, MetaCLGraph. Experimental results show that MetaCLGraph shows better performance compared to both baseline CL methods and developed GCL techniques. The experiments were conducted on various graph datasets including Citeseer, Corafull, Arxiv, and Reddit.